1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/memory_hotplug.c
8 #include <linux/stddef.h>
10 #include <linux/sched/signal.h>
11 #include <linux/swap.h>
12 #include <linux/interrupt.h>
13 #include <linux/pagemap.h>
14 #include <linux/compiler.h>
15 #include <linux/export.h>
16 #include <linux/pagevec.h>
17 #include <linux/writeback.h>
18 #include <linux/slab.h>
19 #include <linux/sysctl.h>
20 #include <linux/cpu.h>
21 #include <linux/memory.h>
22 #include <linux/memremap.h>
23 #include <linux/memory_hotplug.h>
24 #include <linux/vmalloc.h>
25 #include <linux/ioport.h>
26 #include <linux/delay.h>
27 #include <linux/migrate.h>
28 #include <linux/page-isolation.h>
29 #include <linux/pfn.h>
30 #include <linux/suspend.h>
31 #include <linux/mm_inline.h>
32 #include <linux/firmware-map.h>
33 #include <linux/stop_machine.h>
34 #include <linux/hugetlb.h>
35 #include <linux/memblock.h>
36 #include <linux/compaction.h>
37 #include <linux/rmap.h>
38 #include <linux/module.h>
40 #include <asm/tlbflush.h>
47 * memory_hotplug.memmap_on_memory parameter
49 static bool memmap_on_memory __ro_after_init;
50 #ifdef CONFIG_MHP_MEMMAP_ON_MEMORY
51 module_param(memmap_on_memory, bool, 0444);
52 MODULE_PARM_DESC(memmap_on_memory, "Enable memmap on memory for memory hotplug");
56 ONLINE_POLICY_CONTIG_ZONES = 0,
57 ONLINE_POLICY_AUTO_MOVABLE,
60 static const char * const online_policy_to_str[] = {
61 [ONLINE_POLICY_CONTIG_ZONES] = "contig-zones",
62 [ONLINE_POLICY_AUTO_MOVABLE] = "auto-movable",
65 static int set_online_policy(const char *val, const struct kernel_param *kp)
67 int ret = sysfs_match_string(online_policy_to_str, val);
71 *((int *)kp->arg) = ret;
75 static int get_online_policy(char *buffer, const struct kernel_param *kp)
77 return sprintf(buffer, "%s\n", online_policy_to_str[*((int *)kp->arg)]);
81 * memory_hotplug.online_policy: configure online behavior when onlining without
82 * specifying a zone (MMOP_ONLINE)
84 * "contig-zones": keep zone contiguous
85 * "auto-movable": online memory to ZONE_MOVABLE if the configuration
86 * (auto_movable_ratio, auto_movable_numa_aware) allows for it
88 static int online_policy __read_mostly = ONLINE_POLICY_CONTIG_ZONES;
89 static const struct kernel_param_ops online_policy_ops = {
90 .set = set_online_policy,
91 .get = get_online_policy,
93 module_param_cb(online_policy, &online_policy_ops, &online_policy, 0644);
94 MODULE_PARM_DESC(online_policy,
95 "Set the online policy (\"contig-zones\", \"auto-movable\") "
96 "Default: \"contig-zones\"");
99 * memory_hotplug.auto_movable_ratio: specify maximum MOVABLE:KERNEL ratio
101 * The ratio represent an upper limit and the kernel might decide to not
102 * online some memory to ZONE_MOVABLE -- e.g., because hotplugged KERNEL memory
103 * doesn't allow for more MOVABLE memory.
105 static unsigned int auto_movable_ratio __read_mostly = 301;
106 module_param(auto_movable_ratio, uint, 0644);
107 MODULE_PARM_DESC(auto_movable_ratio,
108 "Set the maximum ratio of MOVABLE:KERNEL memory in the system "
109 "in percent for \"auto-movable\" online policy. Default: 301");
112 * memory_hotplug.auto_movable_numa_aware: consider numa node stats
115 static bool auto_movable_numa_aware __read_mostly = true;
116 module_param(auto_movable_numa_aware, bool, 0644);
117 MODULE_PARM_DESC(auto_movable_numa_aware,
118 "Consider numa node stats in addition to global stats in "
119 "\"auto-movable\" online policy. Default: true");
120 #endif /* CONFIG_NUMA */
123 * online_page_callback contains pointer to current page onlining function.
124 * Initially it is generic_online_page(). If it is required it could be
125 * changed by calling set_online_page_callback() for callback registration
126 * and restore_online_page_callback() for generic callback restore.
129 static online_page_callback_t online_page_callback = generic_online_page;
130 static DEFINE_MUTEX(online_page_callback_lock);
132 DEFINE_STATIC_PERCPU_RWSEM(mem_hotplug_lock);
134 void get_online_mems(void)
136 percpu_down_read(&mem_hotplug_lock);
139 void put_online_mems(void)
141 percpu_up_read(&mem_hotplug_lock);
144 bool movable_node_enabled = false;
146 #ifndef CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE
147 int mhp_default_online_type = MMOP_OFFLINE;
149 int mhp_default_online_type = MMOP_ONLINE;
152 static int __init setup_memhp_default_state(char *str)
154 const int online_type = mhp_online_type_from_str(str);
156 if (online_type >= 0)
157 mhp_default_online_type = online_type;
161 __setup("memhp_default_state=", setup_memhp_default_state);
163 void mem_hotplug_begin(void)
166 percpu_down_write(&mem_hotplug_lock);
169 void mem_hotplug_done(void)
171 percpu_up_write(&mem_hotplug_lock);
175 u64 max_mem_size = U64_MAX;
177 /* add this memory to iomem resource */
178 static struct resource *register_memory_resource(u64 start, u64 size,
179 const char *resource_name)
181 struct resource *res;
182 unsigned long flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
184 if (strcmp(resource_name, "System RAM"))
185 flags |= IORESOURCE_SYSRAM_DRIVER_MANAGED;
187 if (!mhp_range_allowed(start, size, true))
188 return ERR_PTR(-E2BIG);
191 * Make sure value parsed from 'mem=' only restricts memory adding
192 * while booting, so that memory hotplug won't be impacted. Please
193 * refer to document of 'mem=' in kernel-parameters.txt for more
196 if (start + size > max_mem_size && system_state < SYSTEM_RUNNING)
197 return ERR_PTR(-E2BIG);
200 * Request ownership of the new memory range. This might be
201 * a child of an existing resource that was present but
202 * not marked as busy.
204 res = __request_region(&iomem_resource, start, size,
205 resource_name, flags);
208 pr_debug("Unable to reserve System RAM region: %016llx->%016llx\n",
209 start, start + size);
210 return ERR_PTR(-EEXIST);
215 static void release_memory_resource(struct resource *res)
219 release_resource(res);
223 static int check_pfn_span(unsigned long pfn, unsigned long nr_pages,
227 * Disallow all operations smaller than a sub-section and only
228 * allow operations smaller than a section for
229 * SPARSEMEM_VMEMMAP. Note that check_hotplug_memory_range()
230 * enforces a larger memory_block_size_bytes() granularity for
231 * memory that will be marked online, so this check should only
232 * fire for direct arch_{add,remove}_memory() users outside of
233 * add_memory_resource().
235 unsigned long min_align;
237 if (IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP))
238 min_align = PAGES_PER_SUBSECTION;
240 min_align = PAGES_PER_SECTION;
241 if (!IS_ALIGNED(pfn, min_align)
242 || !IS_ALIGNED(nr_pages, min_align)) {
243 WARN(1, "Misaligned __%s_pages start: %#lx end: #%lx\n",
244 reason, pfn, pfn + nr_pages - 1);
251 * Return page for the valid pfn only if the page is online. All pfn
252 * walkers which rely on the fully initialized page->flags and others
253 * should use this rather than pfn_valid && pfn_to_page
255 struct page *pfn_to_online_page(unsigned long pfn)
257 unsigned long nr = pfn_to_section_nr(pfn);
258 struct dev_pagemap *pgmap;
259 struct mem_section *ms;
261 if (nr >= NR_MEM_SECTIONS)
264 ms = __nr_to_section(nr);
265 if (!online_section(ms))
269 * Save some code text when online_section() +
270 * pfn_section_valid() are sufficient.
272 if (IS_ENABLED(CONFIG_HAVE_ARCH_PFN_VALID) && !pfn_valid(pfn))
275 if (!pfn_section_valid(ms, pfn))
278 if (!online_device_section(ms))
279 return pfn_to_page(pfn);
282 * Slowpath: when ZONE_DEVICE collides with
283 * ZONE_{NORMAL,MOVABLE} within the same section some pfns in
284 * the section may be 'offline' but 'valid'. Only
285 * get_dev_pagemap() can determine sub-section online status.
287 pgmap = get_dev_pagemap(pfn, NULL);
288 put_dev_pagemap(pgmap);
290 /* The presence of a pgmap indicates ZONE_DEVICE offline pfn */
294 return pfn_to_page(pfn);
296 EXPORT_SYMBOL_GPL(pfn_to_online_page);
299 * Reasonably generic function for adding memory. It is
300 * expected that archs that support memory hotplug will
301 * call this function after deciding the zone to which to
304 int __ref __add_pages(int nid, unsigned long pfn, unsigned long nr_pages,
305 struct mhp_params *params)
307 const unsigned long end_pfn = pfn + nr_pages;
308 unsigned long cur_nr_pages;
310 struct vmem_altmap *altmap = params->altmap;
312 if (WARN_ON_ONCE(!params->pgprot.pgprot))
315 VM_BUG_ON(!mhp_range_allowed(PFN_PHYS(pfn), nr_pages * PAGE_SIZE, false));
319 * Validate altmap is within bounds of the total request
321 if (altmap->base_pfn != pfn
322 || vmem_altmap_offset(altmap) > nr_pages) {
323 pr_warn_once("memory add fail, invalid altmap\n");
329 err = check_pfn_span(pfn, nr_pages, "add");
333 for (; pfn < end_pfn; pfn += cur_nr_pages) {
334 /* Select all remaining pages up to the next section boundary */
335 cur_nr_pages = min(end_pfn - pfn,
336 SECTION_ALIGN_UP(pfn + 1) - pfn);
337 err = sparse_add_section(nid, pfn, cur_nr_pages, altmap);
342 vmemmap_populate_print_last();
346 /* find the smallest valid pfn in the range [start_pfn, end_pfn) */
347 static unsigned long find_smallest_section_pfn(int nid, struct zone *zone,
348 unsigned long start_pfn,
349 unsigned long end_pfn)
351 for (; start_pfn < end_pfn; start_pfn += PAGES_PER_SUBSECTION) {
352 if (unlikely(!pfn_to_online_page(start_pfn)))
355 if (unlikely(pfn_to_nid(start_pfn) != nid))
358 if (zone != page_zone(pfn_to_page(start_pfn)))
367 /* find the biggest valid pfn in the range [start_pfn, end_pfn). */
368 static unsigned long find_biggest_section_pfn(int nid, struct zone *zone,
369 unsigned long start_pfn,
370 unsigned long end_pfn)
374 /* pfn is the end pfn of a memory section. */
376 for (; pfn >= start_pfn; pfn -= PAGES_PER_SUBSECTION) {
377 if (unlikely(!pfn_to_online_page(pfn)))
380 if (unlikely(pfn_to_nid(pfn) != nid))
383 if (zone != page_zone(pfn_to_page(pfn)))
392 static void shrink_zone_span(struct zone *zone, unsigned long start_pfn,
393 unsigned long end_pfn)
396 int nid = zone_to_nid(zone);
398 if (zone->zone_start_pfn == start_pfn) {
400 * If the section is smallest section in the zone, it need
401 * shrink zone->zone_start_pfn and zone->zone_spanned_pages.
402 * In this case, we find second smallest valid mem_section
403 * for shrinking zone.
405 pfn = find_smallest_section_pfn(nid, zone, end_pfn,
408 zone->spanned_pages = zone_end_pfn(zone) - pfn;
409 zone->zone_start_pfn = pfn;
411 zone->zone_start_pfn = 0;
412 zone->spanned_pages = 0;
414 } else if (zone_end_pfn(zone) == end_pfn) {
416 * If the section is biggest section in the zone, it need
417 * shrink zone->spanned_pages.
418 * In this case, we find second biggest valid mem_section for
421 pfn = find_biggest_section_pfn(nid, zone, zone->zone_start_pfn,
424 zone->spanned_pages = pfn - zone->zone_start_pfn + 1;
426 zone->zone_start_pfn = 0;
427 zone->spanned_pages = 0;
432 static void update_pgdat_span(struct pglist_data *pgdat)
434 unsigned long node_start_pfn = 0, node_end_pfn = 0;
437 for (zone = pgdat->node_zones;
438 zone < pgdat->node_zones + MAX_NR_ZONES; zone++) {
439 unsigned long end_pfn = zone_end_pfn(zone);
441 /* No need to lock the zones, they can't change. */
442 if (!zone->spanned_pages)
445 node_start_pfn = zone->zone_start_pfn;
446 node_end_pfn = end_pfn;
450 if (end_pfn > node_end_pfn)
451 node_end_pfn = end_pfn;
452 if (zone->zone_start_pfn < node_start_pfn)
453 node_start_pfn = zone->zone_start_pfn;
456 pgdat->node_start_pfn = node_start_pfn;
457 pgdat->node_spanned_pages = node_end_pfn - node_start_pfn;
460 void __ref remove_pfn_range_from_zone(struct zone *zone,
461 unsigned long start_pfn,
462 unsigned long nr_pages)
464 const unsigned long end_pfn = start_pfn + nr_pages;
465 struct pglist_data *pgdat = zone->zone_pgdat;
466 unsigned long pfn, cur_nr_pages;
468 /* Poison struct pages because they are now uninitialized again. */
469 for (pfn = start_pfn; pfn < end_pfn; pfn += cur_nr_pages) {
472 /* Select all remaining pages up to the next section boundary */
474 min(end_pfn - pfn, SECTION_ALIGN_UP(pfn + 1) - pfn);
475 page_init_poison(pfn_to_page(pfn),
476 sizeof(struct page) * cur_nr_pages);
480 * Zone shrinking code cannot properly deal with ZONE_DEVICE. So
481 * we will not try to shrink the zones - which is okay as
482 * set_zone_contiguous() cannot deal with ZONE_DEVICE either way.
484 if (zone_is_zone_device(zone))
487 clear_zone_contiguous(zone);
489 shrink_zone_span(zone, start_pfn, start_pfn + nr_pages);
490 update_pgdat_span(pgdat);
492 set_zone_contiguous(zone);
495 static void __remove_section(unsigned long pfn, unsigned long nr_pages,
496 unsigned long map_offset,
497 struct vmem_altmap *altmap)
499 struct mem_section *ms = __pfn_to_section(pfn);
501 if (WARN_ON_ONCE(!valid_section(ms)))
504 sparse_remove_section(ms, pfn, nr_pages, map_offset, altmap);
508 * __remove_pages() - remove sections of pages
509 * @pfn: starting pageframe (must be aligned to start of a section)
510 * @nr_pages: number of pages to remove (must be multiple of section size)
511 * @altmap: alternative device page map or %NULL if default memmap is used
513 * Generic helper function to remove section mappings and sysfs entries
514 * for the section of the memory we are removing. Caller needs to make
515 * sure that pages are marked reserved and zones are adjust properly by
516 * calling offline_pages().
518 void __remove_pages(unsigned long pfn, unsigned long nr_pages,
519 struct vmem_altmap *altmap)
521 const unsigned long end_pfn = pfn + nr_pages;
522 unsigned long cur_nr_pages;
523 unsigned long map_offset = 0;
525 map_offset = vmem_altmap_offset(altmap);
527 if (check_pfn_span(pfn, nr_pages, "remove"))
530 for (; pfn < end_pfn; pfn += cur_nr_pages) {
532 /* Select all remaining pages up to the next section boundary */
533 cur_nr_pages = min(end_pfn - pfn,
534 SECTION_ALIGN_UP(pfn + 1) - pfn);
535 __remove_section(pfn, cur_nr_pages, map_offset, altmap);
540 int set_online_page_callback(online_page_callback_t callback)
545 mutex_lock(&online_page_callback_lock);
547 if (online_page_callback == generic_online_page) {
548 online_page_callback = callback;
552 mutex_unlock(&online_page_callback_lock);
557 EXPORT_SYMBOL_GPL(set_online_page_callback);
559 int restore_online_page_callback(online_page_callback_t callback)
564 mutex_lock(&online_page_callback_lock);
566 if (online_page_callback == callback) {
567 online_page_callback = generic_online_page;
571 mutex_unlock(&online_page_callback_lock);
576 EXPORT_SYMBOL_GPL(restore_online_page_callback);
578 void generic_online_page(struct page *page, unsigned int order)
581 * Freeing the page with debug_pagealloc enabled will try to unmap it,
582 * so we should map it first. This is better than introducing a special
583 * case in page freeing fast path.
585 debug_pagealloc_map_pages(page, 1 << order);
586 __free_pages_core(page, order);
587 totalram_pages_add(1UL << order);
589 EXPORT_SYMBOL_GPL(generic_online_page);
591 static void online_pages_range(unsigned long start_pfn, unsigned long nr_pages)
593 const unsigned long end_pfn = start_pfn + nr_pages;
597 * Online the pages in MAX_ORDER - 1 aligned chunks. The callback might
598 * decide to not expose all pages to the buddy (e.g., expose them
599 * later). We account all pages as being online and belonging to this
601 * When using memmap_on_memory, the range might not be aligned to
602 * MAX_ORDER_NR_PAGES - 1, but pageblock aligned. __ffs() will detect
603 * this and the first chunk to online will be pageblock_nr_pages.
605 for (pfn = start_pfn; pfn < end_pfn;) {
606 int order = min(MAX_ORDER - 1UL, __ffs(pfn));
608 (*online_page_callback)(pfn_to_page(pfn), order);
609 pfn += (1UL << order);
612 /* mark all involved sections as online */
613 online_mem_sections(start_pfn, end_pfn);
616 /* check which state of node_states will be changed when online memory */
617 static void node_states_check_changes_online(unsigned long nr_pages,
618 struct zone *zone, struct memory_notify *arg)
620 int nid = zone_to_nid(zone);
622 arg->status_change_nid = NUMA_NO_NODE;
623 arg->status_change_nid_normal = NUMA_NO_NODE;
625 if (!node_state(nid, N_MEMORY))
626 arg->status_change_nid = nid;
627 if (zone_idx(zone) <= ZONE_NORMAL && !node_state(nid, N_NORMAL_MEMORY))
628 arg->status_change_nid_normal = nid;
631 static void node_states_set_node(int node, struct memory_notify *arg)
633 if (arg->status_change_nid_normal >= 0)
634 node_set_state(node, N_NORMAL_MEMORY);
636 if (arg->status_change_nid >= 0)
637 node_set_state(node, N_MEMORY);
640 static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn,
641 unsigned long nr_pages)
643 unsigned long old_end_pfn = zone_end_pfn(zone);
645 if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn)
646 zone->zone_start_pfn = start_pfn;
648 zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn;
651 static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn,
652 unsigned long nr_pages)
654 unsigned long old_end_pfn = pgdat_end_pfn(pgdat);
656 if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn)
657 pgdat->node_start_pfn = start_pfn;
659 pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn;
663 static void section_taint_zone_device(unsigned long pfn)
665 struct mem_section *ms = __pfn_to_section(pfn);
667 ms->section_mem_map |= SECTION_TAINT_ZONE_DEVICE;
671 * Associate the pfn range with the given zone, initializing the memmaps
672 * and resizing the pgdat/zone data to span the added pages. After this
673 * call, all affected pages are PG_reserved.
675 * All aligned pageblocks are initialized to the specified migratetype
676 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
677 * zone stats (e.g., nr_isolate_pageblock) are touched.
679 void __ref move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn,
680 unsigned long nr_pages,
681 struct vmem_altmap *altmap, int migratetype)
683 struct pglist_data *pgdat = zone->zone_pgdat;
684 int nid = pgdat->node_id;
686 clear_zone_contiguous(zone);
688 if (zone_is_empty(zone))
689 init_currently_empty_zone(zone, start_pfn, nr_pages);
690 resize_zone_range(zone, start_pfn, nr_pages);
691 resize_pgdat_range(pgdat, start_pfn, nr_pages);
694 * Subsection population requires care in pfn_to_online_page().
695 * Set the taint to enable the slow path detection of
696 * ZONE_DEVICE pages in an otherwise ZONE_{NORMAL,MOVABLE}
699 if (zone_is_zone_device(zone)) {
700 if (!IS_ALIGNED(start_pfn, PAGES_PER_SECTION))
701 section_taint_zone_device(start_pfn);
702 if (!IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION))
703 section_taint_zone_device(start_pfn + nr_pages);
707 * TODO now we have a visible range of pages which are not associated
708 * with their zone properly. Not nice but set_pfnblock_flags_mask
709 * expects the zone spans the pfn range. All the pages in the range
710 * are reserved so nobody should be touching them so we should be safe
712 memmap_init_range(nr_pages, nid, zone_idx(zone), start_pfn, 0,
713 MEMINIT_HOTPLUG, altmap, migratetype);
715 set_zone_contiguous(zone);
718 struct auto_movable_stats {
719 unsigned long kernel_early_pages;
720 unsigned long movable_pages;
723 static void auto_movable_stats_account_zone(struct auto_movable_stats *stats,
726 if (zone_idx(zone) == ZONE_MOVABLE) {
727 stats->movable_pages += zone->present_pages;
729 stats->kernel_early_pages += zone->present_early_pages;
732 * CMA pages (never on hotplugged memory) behave like
735 stats->movable_pages += zone->cma_pages;
736 stats->kernel_early_pages -= zone->cma_pages;
737 #endif /* CONFIG_CMA */
740 struct auto_movable_group_stats {
741 unsigned long movable_pages;
742 unsigned long req_kernel_early_pages;
745 static int auto_movable_stats_account_group(struct memory_group *group,
748 const int ratio = READ_ONCE(auto_movable_ratio);
749 struct auto_movable_group_stats *stats = arg;
753 * We don't support modifying the config while the auto-movable online
754 * policy is already enabled. Just avoid the division by zero below.
760 * Calculate how many early kernel pages this group requires to
761 * satisfy the configured zone ratio.
763 pages = group->present_movable_pages * 100 / ratio;
764 pages -= group->present_kernel_pages;
767 stats->req_kernel_early_pages += pages;
768 stats->movable_pages += group->present_movable_pages;
772 static bool auto_movable_can_online_movable(int nid, struct memory_group *group,
773 unsigned long nr_pages)
775 unsigned long kernel_early_pages, movable_pages;
776 struct auto_movable_group_stats group_stats = {};
777 struct auto_movable_stats stats = {};
778 pg_data_t *pgdat = NODE_DATA(nid);
782 /* Walk all relevant zones and collect MOVABLE vs. KERNEL stats. */
783 if (nid == NUMA_NO_NODE) {
784 /* TODO: cache values */
785 for_each_populated_zone(zone)
786 auto_movable_stats_account_zone(&stats, zone);
788 for (i = 0; i < MAX_NR_ZONES; i++) {
789 zone = pgdat->node_zones + i;
790 if (populated_zone(zone))
791 auto_movable_stats_account_zone(&stats, zone);
795 kernel_early_pages = stats.kernel_early_pages;
796 movable_pages = stats.movable_pages;
799 * Kernel memory inside dynamic memory group allows for more MOVABLE
800 * memory within the same group. Remove the effect of all but the
801 * current group from the stats.
803 walk_dynamic_memory_groups(nid, auto_movable_stats_account_group,
804 group, &group_stats);
805 if (kernel_early_pages <= group_stats.req_kernel_early_pages)
807 kernel_early_pages -= group_stats.req_kernel_early_pages;
808 movable_pages -= group_stats.movable_pages;
810 if (group && group->is_dynamic)
811 kernel_early_pages += group->present_kernel_pages;
814 * Test if we could online the given number of pages to ZONE_MOVABLE
815 * and still stay in the configured ratio.
817 movable_pages += nr_pages;
818 return movable_pages <= (auto_movable_ratio * kernel_early_pages) / 100;
822 * Returns a default kernel memory zone for the given pfn range.
823 * If no kernel zone covers this pfn range it will automatically go
824 * to the ZONE_NORMAL.
826 static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn,
827 unsigned long nr_pages)
829 struct pglist_data *pgdat = NODE_DATA(nid);
832 for (zid = 0; zid <= ZONE_NORMAL; zid++) {
833 struct zone *zone = &pgdat->node_zones[zid];
835 if (zone_intersects(zone, start_pfn, nr_pages))
839 return &pgdat->node_zones[ZONE_NORMAL];
843 * Determine to which zone to online memory dynamically based on user
844 * configuration and system stats. We care about the following ratio:
848 * Whereby MOVABLE is memory in ZONE_MOVABLE and KERNEL is memory in
849 * one of the kernel zones. CMA pages inside one of the kernel zones really
850 * behaves like ZONE_MOVABLE, so we treat them accordingly.
852 * We don't allow for hotplugged memory in a KERNEL zone to increase the
853 * amount of MOVABLE memory we can have, so we end up with:
855 * MOVABLE : KERNEL_EARLY
857 * Whereby KERNEL_EARLY is memory in one of the kernel zones, available sinze
858 * boot. We base our calculation on KERNEL_EARLY internally, because:
860 * a) Hotplugged memory in one of the kernel zones can sometimes still get
861 * hotunplugged, especially when hot(un)plugging individual memory blocks.
862 * There is no coordination across memory devices, therefore "automatic"
863 * hotunplugging, as implemented in hypervisors, could result in zone
865 * b) Early/boot memory in one of the kernel zones can usually not get
866 * hotunplugged again (e.g., no firmware interface to unplug, fragmented
867 * with unmovable allocations). While there are corner cases where it might
868 * still work, it is barely relevant in practice.
870 * Exceptions are dynamic memory groups, which allow for more MOVABLE
871 * memory within the same memory group -- because in that case, there is
872 * coordination within the single memory device managed by a single driver.
874 * We rely on "present pages" instead of "managed pages", as the latter is
875 * highly unreliable and dynamic in virtualized environments, and does not
876 * consider boot time allocations. For example, memory ballooning adjusts the
877 * managed pages when inflating/deflating the balloon, and balloon compaction
878 * can even migrate inflated pages between zones.
880 * Using "present pages" is better but some things to keep in mind are:
882 * a) Some memblock allocations, such as for the crashkernel area, are
883 * effectively unused by the kernel, yet they account to "present pages".
884 * Fortunately, these allocations are comparatively small in relevant setups
885 * (e.g., fraction of system memory).
886 * b) Some hotplugged memory blocks in virtualized environments, esecially
887 * hotplugged by virtio-mem, look like they are completely present, however,
888 * only parts of the memory block are actually currently usable.
889 * "present pages" is an upper limit that can get reached at runtime. As
890 * we base our calculations on KERNEL_EARLY, this is not an issue.
892 static struct zone *auto_movable_zone_for_pfn(int nid,
893 struct memory_group *group,
895 unsigned long nr_pages)
897 unsigned long online_pages = 0, max_pages, end_pfn;
900 if (!auto_movable_ratio)
903 if (group && !group->is_dynamic) {
904 max_pages = group->s.max_pages;
905 online_pages = group->present_movable_pages;
907 /* If anything is !MOVABLE online the rest !MOVABLE. */
908 if (group->present_kernel_pages)
910 } else if (!group || group->d.unit_pages == nr_pages) {
911 max_pages = nr_pages;
913 max_pages = group->d.unit_pages;
915 * Take a look at all online sections in the current unit.
916 * We can safely assume that all pages within a section belong
917 * to the same zone, because dynamic memory groups only deal
918 * with hotplugged memory.
920 pfn = ALIGN_DOWN(pfn, group->d.unit_pages);
921 end_pfn = pfn + group->d.unit_pages;
922 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
923 page = pfn_to_online_page(pfn);
926 /* If anything is !MOVABLE online the rest !MOVABLE. */
927 if (page_zonenum(page) != ZONE_MOVABLE)
929 online_pages += PAGES_PER_SECTION;
934 * Online MOVABLE if we could *currently* online all remaining parts
935 * MOVABLE. We expect to (add+) online them immediately next, so if
936 * nobody interferes, all will be MOVABLE if possible.
938 nr_pages = max_pages - online_pages;
939 if (!auto_movable_can_online_movable(NUMA_NO_NODE, group, nr_pages))
943 if (auto_movable_numa_aware &&
944 !auto_movable_can_online_movable(nid, group, nr_pages))
946 #endif /* CONFIG_NUMA */
948 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
950 return default_kernel_zone_for_pfn(nid, pfn, nr_pages);
953 static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn,
954 unsigned long nr_pages)
956 struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn,
958 struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
959 bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages);
960 bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages);
963 * We inherit the existing zone in a simple case where zones do not
964 * overlap in the given range
966 if (in_kernel ^ in_movable)
967 return (in_kernel) ? kernel_zone : movable_zone;
970 * If the range doesn't belong to any zone or two zones overlap in the
971 * given range then we use movable zone only if movable_node is
972 * enabled because we always online to a kernel zone by default.
974 return movable_node_enabled ? movable_zone : kernel_zone;
977 struct zone *zone_for_pfn_range(int online_type, int nid,
978 struct memory_group *group, unsigned long start_pfn,
979 unsigned long nr_pages)
981 if (online_type == MMOP_ONLINE_KERNEL)
982 return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages);
984 if (online_type == MMOP_ONLINE_MOVABLE)
985 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
987 if (online_policy == ONLINE_POLICY_AUTO_MOVABLE)
988 return auto_movable_zone_for_pfn(nid, group, start_pfn, nr_pages);
990 return default_zone_for_pfn(nid, start_pfn, nr_pages);
994 * This function should only be called by memory_block_{online,offline},
995 * and {online,offline}_pages.
997 void adjust_present_page_count(struct page *page, struct memory_group *group,
1000 struct zone *zone = page_zone(page);
1001 const bool movable = zone_idx(zone) == ZONE_MOVABLE;
1004 * We only support onlining/offlining/adding/removing of complete
1005 * memory blocks; therefore, either all is either early or hotplugged.
1007 if (early_section(__pfn_to_section(page_to_pfn(page))))
1008 zone->present_early_pages += nr_pages;
1009 zone->present_pages += nr_pages;
1010 zone->zone_pgdat->node_present_pages += nr_pages;
1012 if (group && movable)
1013 group->present_movable_pages += nr_pages;
1014 else if (group && !movable)
1015 group->present_kernel_pages += nr_pages;
1018 int mhp_init_memmap_on_memory(unsigned long pfn, unsigned long nr_pages,
1021 unsigned long end_pfn = pfn + nr_pages;
1024 ret = kasan_add_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1028 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_UNMOVABLE);
1031 * It might be that the vmemmap_pages fully span sections. If that is
1032 * the case, mark those sections online here as otherwise they will be
1035 if (nr_pages >= PAGES_PER_SECTION)
1036 online_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1041 void mhp_deinit_memmap_on_memory(unsigned long pfn, unsigned long nr_pages)
1043 unsigned long end_pfn = pfn + nr_pages;
1046 * It might be that the vmemmap_pages fully span sections. If that is
1047 * the case, mark those sections offline here as otherwise they will be
1050 if (nr_pages >= PAGES_PER_SECTION)
1051 offline_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1054 * The pages associated with this vmemmap have been offlined, so
1055 * we can reset its state here.
1057 remove_pfn_range_from_zone(page_zone(pfn_to_page(pfn)), pfn, nr_pages);
1058 kasan_remove_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1061 int __ref online_pages(unsigned long pfn, unsigned long nr_pages,
1062 struct zone *zone, struct memory_group *group)
1064 unsigned long flags;
1065 int need_zonelists_rebuild = 0;
1066 const int nid = zone_to_nid(zone);
1068 struct memory_notify arg;
1071 * {on,off}lining is constrained to full memory sections (or more
1072 * precisely to memory blocks from the user space POV).
1073 * memmap_on_memory is an exception because it reserves initial part
1074 * of the physical memory space for vmemmaps. That space is pageblock
1077 if (WARN_ON_ONCE(!nr_pages ||
1078 !IS_ALIGNED(pfn, pageblock_nr_pages) ||
1079 !IS_ALIGNED(pfn + nr_pages, PAGES_PER_SECTION)))
1082 mem_hotplug_begin();
1084 /* associate pfn range with the zone */
1085 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_ISOLATE);
1087 arg.start_pfn = pfn;
1088 arg.nr_pages = nr_pages;
1089 node_states_check_changes_online(nr_pages, zone, &arg);
1091 ret = memory_notify(MEM_GOING_ONLINE, &arg);
1092 ret = notifier_to_errno(ret);
1094 goto failed_addition;
1097 * Fixup the number of isolated pageblocks before marking the sections
1098 * onlining, such that undo_isolate_page_range() works correctly.
1100 spin_lock_irqsave(&zone->lock, flags);
1101 zone->nr_isolate_pageblock += nr_pages / pageblock_nr_pages;
1102 spin_unlock_irqrestore(&zone->lock, flags);
1105 * If this zone is not populated, then it is not in zonelist.
1106 * This means the page allocator ignores this zone.
1107 * So, zonelist must be updated after online.
1109 if (!populated_zone(zone)) {
1110 need_zonelists_rebuild = 1;
1111 setup_zone_pageset(zone);
1114 online_pages_range(pfn, nr_pages);
1115 adjust_present_page_count(pfn_to_page(pfn), group, nr_pages);
1117 node_states_set_node(nid, &arg);
1118 if (need_zonelists_rebuild)
1119 build_all_zonelists(NULL);
1121 /* Basic onlining is complete, allow allocation of onlined pages. */
1122 undo_isolate_page_range(pfn, pfn + nr_pages, MIGRATE_MOVABLE);
1125 * Freshly onlined pages aren't shuffled (e.g., all pages are placed to
1126 * the tail of the freelist when undoing isolation). Shuffle the whole
1127 * zone to make sure the just onlined pages are properly distributed
1128 * across the whole freelist - to create an initial shuffle.
1132 /* reinitialise watermarks and update pcp limits */
1133 init_per_zone_wmark_min();
1138 writeback_set_ratelimit();
1140 memory_notify(MEM_ONLINE, &arg);
1145 pr_debug("online_pages [mem %#010llx-%#010llx] failed\n",
1146 (unsigned long long) pfn << PAGE_SHIFT,
1147 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1);
1148 memory_notify(MEM_CANCEL_ONLINE, &arg);
1149 remove_pfn_range_from_zone(zone, pfn, nr_pages);
1154 static void reset_node_present_pages(pg_data_t *pgdat)
1158 for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++)
1159 z->present_pages = 0;
1161 pgdat->node_present_pages = 0;
1164 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
1165 static pg_data_t __ref *hotadd_new_pgdat(int nid)
1167 struct pglist_data *pgdat;
1169 pgdat = NODE_DATA(nid);
1171 pgdat = arch_alloc_nodedata(nid);
1175 pgdat->per_cpu_nodestats =
1176 alloc_percpu(struct per_cpu_nodestat);
1177 arch_refresh_nodedata(nid, pgdat);
1181 * Reset the nr_zones, order and highest_zoneidx before reuse.
1182 * Note that kswapd will init kswapd_highest_zoneidx properly
1183 * when it starts in the near future.
1185 pgdat->nr_zones = 0;
1186 pgdat->kswapd_order = 0;
1187 pgdat->kswapd_highest_zoneidx = 0;
1188 for_each_online_cpu(cpu) {
1189 struct per_cpu_nodestat *p;
1191 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
1192 memset(p, 0, sizeof(*p));
1196 /* we can use NODE_DATA(nid) from here */
1197 pgdat->node_id = nid;
1198 pgdat->node_start_pfn = 0;
1200 /* init node's zones as empty zones, we don't have any present pages.*/
1201 free_area_init_core_hotplug(nid);
1204 * The node we allocated has no zone fallback lists. For avoiding
1205 * to access not-initialized zonelist, build here.
1207 build_all_zonelists(pgdat);
1210 * When memory is hot-added, all the memory is in offline state. So
1211 * clear all zones' present_pages because they will be updated in
1212 * online_pages() and offline_pages().
1214 reset_node_managed_pages(pgdat);
1215 reset_node_present_pages(pgdat);
1220 static void rollback_node_hotadd(int nid)
1222 pg_data_t *pgdat = NODE_DATA(nid);
1224 arch_refresh_nodedata(nid, NULL);
1225 free_percpu(pgdat->per_cpu_nodestats);
1226 arch_free_nodedata(pgdat);
1231 * __try_online_node - online a node if offlined
1233 * @set_node_online: Whether we want to online the node
1234 * called by cpu_up() to online a node without onlined memory.
1237 * 1 -> a new node has been allocated
1238 * 0 -> the node is already online
1239 * -ENOMEM -> the node could not be allocated
1241 static int __try_online_node(int nid, bool set_node_online)
1246 if (node_online(nid))
1249 pgdat = hotadd_new_pgdat(nid);
1251 pr_err("Cannot online node %d due to NULL pgdat\n", nid);
1256 if (set_node_online) {
1257 node_set_online(nid);
1258 ret = register_one_node(nid);
1266 * Users of this function always want to online/register the node
1268 int try_online_node(int nid)
1272 mem_hotplug_begin();
1273 ret = __try_online_node(nid, true);
1278 static int check_hotplug_memory_range(u64 start, u64 size)
1280 /* memory range must be block size aligned */
1281 if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) ||
1282 !IS_ALIGNED(size, memory_block_size_bytes())) {
1283 pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx",
1284 memory_block_size_bytes(), start, size);
1291 static int online_memory_block(struct memory_block *mem, void *arg)
1293 mem->online_type = mhp_default_online_type;
1294 return device_online(&mem->dev);
1297 bool mhp_supports_memmap_on_memory(unsigned long size)
1299 unsigned long nr_vmemmap_pages = size / PAGE_SIZE;
1300 unsigned long vmemmap_size = nr_vmemmap_pages * sizeof(struct page);
1301 unsigned long remaining_size = size - vmemmap_size;
1304 * Besides having arch support and the feature enabled at runtime, we
1305 * need a few more assumptions to hold true:
1307 * a) We span a single memory block: memory onlining/offlinin;g happens
1308 * in memory block granularity. We don't want the vmemmap of online
1309 * memory blocks to reside on offline memory blocks. In the future,
1310 * we might want to support variable-sized memory blocks to make the
1311 * feature more versatile.
1313 * b) The vmemmap pages span complete PMDs: We don't want vmemmap code
1314 * to populate memory from the altmap for unrelated parts (i.e.,
1315 * other memory blocks)
1317 * c) The vmemmap pages (and thereby the pages that will be exposed to
1318 * the buddy) have to cover full pageblocks: memory onlining/offlining
1319 * code requires applicable ranges to be page-aligned, for example, to
1320 * set the migratetypes properly.
1322 * TODO: Although we have a check here to make sure that vmemmap pages
1323 * fully populate a PMD, it is not the right place to check for
1324 * this. A much better solution involves improving vmemmap code
1325 * to fallback to base pages when trying to populate vmemmap using
1326 * altmap as an alternative source of memory, and we do not exactly
1327 * populate a single PMD.
1329 return memmap_on_memory &&
1330 !hugetlb_free_vmemmap_enabled &&
1331 IS_ENABLED(CONFIG_MHP_MEMMAP_ON_MEMORY) &&
1332 size == memory_block_size_bytes() &&
1333 IS_ALIGNED(vmemmap_size, PMD_SIZE) &&
1334 IS_ALIGNED(remaining_size, (pageblock_nr_pages << PAGE_SHIFT));
1338 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1339 * and online/offline operations (triggered e.g. by sysfs).
1341 * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG
1343 int __ref add_memory_resource(int nid, struct resource *res, mhp_t mhp_flags)
1345 struct mhp_params params = { .pgprot = pgprot_mhp(PAGE_KERNEL) };
1346 enum memblock_flags memblock_flags = MEMBLOCK_NONE;
1347 struct vmem_altmap mhp_altmap = {};
1348 struct memory_group *group = NULL;
1350 bool new_node = false;
1354 size = resource_size(res);
1356 ret = check_hotplug_memory_range(start, size);
1360 if (mhp_flags & MHP_NID_IS_MGID) {
1361 group = memory_group_find_by_id(nid);
1367 if (!node_possible(nid)) {
1368 WARN(1, "node %d was absent from the node_possible_map\n", nid);
1372 mem_hotplug_begin();
1374 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
1375 if (res->flags & IORESOURCE_SYSRAM_DRIVER_MANAGED)
1376 memblock_flags = MEMBLOCK_DRIVER_MANAGED;
1377 ret = memblock_add_node(start, size, nid, memblock_flags);
1379 goto error_mem_hotplug_end;
1382 ret = __try_online_node(nid, false);
1388 * Self hosted memmap array
1390 if (mhp_flags & MHP_MEMMAP_ON_MEMORY) {
1391 if (!mhp_supports_memmap_on_memory(size)) {
1395 mhp_altmap.free = PHYS_PFN(size);
1396 mhp_altmap.base_pfn = PHYS_PFN(start);
1397 params.altmap = &mhp_altmap;
1400 /* call arch's memory hotadd */
1401 ret = arch_add_memory(nid, start, size, ¶ms);
1405 /* create memory block devices after memory was added */
1406 ret = create_memory_block_devices(start, size, mhp_altmap.alloc,
1409 arch_remove_memory(start, size, NULL);
1414 /* If sysfs file of new node can't be created, cpu on the node
1415 * can't be hot-added. There is no rollback way now.
1416 * So, check by BUG_ON() to catch it reluctantly..
1417 * We online node here. We can't roll back from here.
1419 node_set_online(nid);
1420 ret = __register_one_node(nid);
1424 /* link memory sections under this node.*/
1425 link_mem_sections(nid, PFN_DOWN(start), PFN_UP(start + size - 1),
1428 /* create new memmap entry */
1429 if (!strcmp(res->name, "System RAM"))
1430 firmware_map_add_hotplug(start, start + size, "System RAM");
1432 /* device_online() will take the lock when calling online_pages() */
1436 * In case we're allowed to merge the resource, flag it and trigger
1437 * merging now that adding succeeded.
1439 if (mhp_flags & MHP_MERGE_RESOURCE)
1440 merge_system_ram_resource(res);
1442 /* online pages if requested */
1443 if (mhp_default_online_type != MMOP_OFFLINE)
1444 walk_memory_blocks(start, size, NULL, online_memory_block);
1448 /* rollback pgdat allocation and others */
1450 rollback_node_hotadd(nid);
1451 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
1452 memblock_remove(start, size);
1453 error_mem_hotplug_end:
1458 /* requires device_hotplug_lock, see add_memory_resource() */
1459 int __ref __add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1461 struct resource *res;
1464 res = register_memory_resource(start, size, "System RAM");
1466 return PTR_ERR(res);
1468 ret = add_memory_resource(nid, res, mhp_flags);
1470 release_memory_resource(res);
1474 int add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1478 lock_device_hotplug();
1479 rc = __add_memory(nid, start, size, mhp_flags);
1480 unlock_device_hotplug();
1484 EXPORT_SYMBOL_GPL(add_memory);
1487 * Add special, driver-managed memory to the system as system RAM. Such
1488 * memory is not exposed via the raw firmware-provided memmap as system
1489 * RAM, instead, it is detected and added by a driver - during cold boot,
1490 * after a reboot, and after kexec.
1492 * Reasons why this memory should not be used for the initial memmap of a
1493 * kexec kernel or for placing kexec images:
1494 * - The booting kernel is in charge of determining how this memory will be
1495 * used (e.g., use persistent memory as system RAM)
1496 * - Coordination with a hypervisor is required before this memory
1497 * can be used (e.g., inaccessible parts).
1499 * For this memory, no entries in /sys/firmware/memmap ("raw firmware-provided
1500 * memory map") are created. Also, the created memory resource is flagged
1501 * with IORESOURCE_SYSRAM_DRIVER_MANAGED, so in-kernel users can special-case
1502 * this memory as well (esp., not place kexec images onto it).
1504 * The resource_name (visible via /proc/iomem) has to have the format
1505 * "System RAM ($DRIVER)".
1507 int add_memory_driver_managed(int nid, u64 start, u64 size,
1508 const char *resource_name, mhp_t mhp_flags)
1510 struct resource *res;
1513 if (!resource_name ||
1514 strstr(resource_name, "System RAM (") != resource_name ||
1515 resource_name[strlen(resource_name) - 1] != ')')
1518 lock_device_hotplug();
1520 res = register_memory_resource(start, size, resource_name);
1526 rc = add_memory_resource(nid, res, mhp_flags);
1528 release_memory_resource(res);
1531 unlock_device_hotplug();
1534 EXPORT_SYMBOL_GPL(add_memory_driver_managed);
1537 * Platforms should define arch_get_mappable_range() that provides
1538 * maximum possible addressable physical memory range for which the
1539 * linear mapping could be created. The platform returned address
1540 * range must adhere to these following semantics.
1542 * - range.start <= range.end
1543 * - Range includes both end points [range.start..range.end]
1545 * There is also a fallback definition provided here, allowing the
1546 * entire possible physical address range in case any platform does
1547 * not define arch_get_mappable_range().
1549 struct range __weak arch_get_mappable_range(void)
1551 struct range mhp_range = {
1558 struct range mhp_get_pluggable_range(bool need_mapping)
1560 const u64 max_phys = (1ULL << MAX_PHYSMEM_BITS) - 1;
1561 struct range mhp_range;
1564 mhp_range = arch_get_mappable_range();
1565 if (mhp_range.start > max_phys) {
1566 mhp_range.start = 0;
1569 mhp_range.end = min_t(u64, mhp_range.end, max_phys);
1571 mhp_range.start = 0;
1572 mhp_range.end = max_phys;
1576 EXPORT_SYMBOL_GPL(mhp_get_pluggable_range);
1578 bool mhp_range_allowed(u64 start, u64 size, bool need_mapping)
1580 struct range mhp_range = mhp_get_pluggable_range(need_mapping);
1581 u64 end = start + size;
1583 if (start < end && start >= mhp_range.start && (end - 1) <= mhp_range.end)
1586 pr_warn("Hotplug memory [%#llx-%#llx] exceeds maximum addressable range [%#llx-%#llx]\n",
1587 start, end, mhp_range.start, mhp_range.end);
1591 #ifdef CONFIG_MEMORY_HOTREMOVE
1593 * Confirm all pages in a range [start, end) belong to the same zone (skipping
1594 * memory holes). When true, return the zone.
1596 struct zone *test_pages_in_a_zone(unsigned long start_pfn,
1597 unsigned long end_pfn)
1599 unsigned long pfn, sec_end_pfn;
1600 struct zone *zone = NULL;
1603 for (pfn = start_pfn, sec_end_pfn = SECTION_ALIGN_UP(start_pfn + 1);
1605 pfn = sec_end_pfn, sec_end_pfn += PAGES_PER_SECTION) {
1606 /* Make sure the memory section is present first */
1607 if (!present_section_nr(pfn_to_section_nr(pfn)))
1609 for (; pfn < sec_end_pfn && pfn < end_pfn;
1610 pfn += MAX_ORDER_NR_PAGES) {
1611 /* Check if we got outside of the zone */
1612 if (zone && !zone_spans_pfn(zone, pfn))
1614 page = pfn_to_page(pfn);
1615 if (zone && page_zone(page) != zone)
1617 zone = page_zone(page);
1625 * Scan pfn range [start,end) to find movable/migratable pages (LRU pages,
1626 * non-lru movable pages and hugepages). Will skip over most unmovable
1627 * pages (esp., pages that can be skipped when offlining), but bail out on
1628 * definitely unmovable pages.
1631 * 0 in case a movable page is found and movable_pfn was updated.
1632 * -ENOENT in case no movable page was found.
1633 * -EBUSY in case a definitely unmovable page was found.
1635 static int scan_movable_pages(unsigned long start, unsigned long end,
1636 unsigned long *movable_pfn)
1640 for (pfn = start; pfn < end; pfn++) {
1641 struct page *page, *head;
1644 if (!pfn_valid(pfn))
1646 page = pfn_to_page(pfn);
1649 if (__PageMovable(page))
1653 * PageOffline() pages that are not marked __PageMovable() and
1654 * have a reference count > 0 (after MEM_GOING_OFFLINE) are
1655 * definitely unmovable. If their reference count would be 0,
1656 * they could at least be skipped when offlining memory.
1658 if (PageOffline(page) && page_count(page))
1661 if (!PageHuge(page))
1663 head = compound_head(page);
1665 * This test is racy as we hold no reference or lock. The
1666 * hugetlb page could have been free'ed and head is no longer
1667 * a hugetlb page before the following check. In such unlikely
1668 * cases false positives and negatives are possible. Calling
1669 * code must deal with these scenarios.
1671 if (HPageMigratable(head))
1673 skip = compound_nr(head) - (page - head);
1683 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
1686 struct page *page, *head;
1689 static DEFINE_RATELIMIT_STATE(migrate_rs, DEFAULT_RATELIMIT_INTERVAL,
1690 DEFAULT_RATELIMIT_BURST);
1692 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1693 if (!pfn_valid(pfn))
1695 page = pfn_to_page(pfn);
1696 head = compound_head(page);
1698 if (PageHuge(page)) {
1699 pfn = page_to_pfn(head) + compound_nr(head) - 1;
1700 isolate_huge_page(head, &source);
1702 } else if (PageTransHuge(page))
1703 pfn = page_to_pfn(head) + thp_nr_pages(page) - 1;
1706 * HWPoison pages have elevated reference counts so the migration would
1707 * fail on them. It also doesn't make any sense to migrate them in the
1708 * first place. Still try to unmap such a page in case it is still mapped
1709 * (e.g. current hwpoison implementation doesn't unmap KSM pages but keep
1710 * the unmap as the catch all safety net).
1712 if (PageHWPoison(page)) {
1713 if (WARN_ON(PageLRU(page)))
1714 isolate_lru_page(page);
1715 if (page_mapped(page))
1716 try_to_unmap(page, TTU_IGNORE_MLOCK);
1720 if (!get_page_unless_zero(page))
1723 * We can skip free pages. And we can deal with pages on
1724 * LRU and non-lru movable pages.
1727 ret = isolate_lru_page(page);
1729 ret = isolate_movable_page(page, ISOLATE_UNEVICTABLE);
1730 if (!ret) { /* Success */
1731 list_add_tail(&page->lru, &source);
1732 if (!__PageMovable(page))
1733 inc_node_page_state(page, NR_ISOLATED_ANON +
1734 page_is_file_lru(page));
1737 if (__ratelimit(&migrate_rs)) {
1738 pr_warn("failed to isolate pfn %lx\n", pfn);
1739 dump_page(page, "isolation failed");
1744 if (!list_empty(&source)) {
1745 nodemask_t nmask = node_states[N_MEMORY];
1746 struct migration_target_control mtc = {
1748 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
1752 * We have checked that migration range is on a single zone so
1753 * we can use the nid of the first page to all the others.
1755 mtc.nid = page_to_nid(list_first_entry(&source, struct page, lru));
1758 * try to allocate from a different node but reuse this node
1759 * if there are no other online nodes to be used (e.g. we are
1760 * offlining a part of the only existing node)
1762 node_clear(mtc.nid, nmask);
1763 if (nodes_empty(nmask))
1764 node_set(mtc.nid, nmask);
1765 ret = migrate_pages(&source, alloc_migration_target, NULL,
1766 (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_HOTPLUG, NULL);
1768 list_for_each_entry(page, &source, lru) {
1769 if (__ratelimit(&migrate_rs)) {
1770 pr_warn("migrating pfn %lx failed ret:%d\n",
1771 page_to_pfn(page), ret);
1772 dump_page(page, "migration failure");
1775 putback_movable_pages(&source);
1782 static int __init cmdline_parse_movable_node(char *p)
1784 movable_node_enabled = true;
1787 early_param("movable_node", cmdline_parse_movable_node);
1789 /* check which state of node_states will be changed when offline memory */
1790 static void node_states_check_changes_offline(unsigned long nr_pages,
1791 struct zone *zone, struct memory_notify *arg)
1793 struct pglist_data *pgdat = zone->zone_pgdat;
1794 unsigned long present_pages = 0;
1797 arg->status_change_nid = NUMA_NO_NODE;
1798 arg->status_change_nid_normal = NUMA_NO_NODE;
1801 * Check whether node_states[N_NORMAL_MEMORY] will be changed.
1802 * If the memory to be offline is within the range
1803 * [0..ZONE_NORMAL], and it is the last present memory there,
1804 * the zones in that range will become empty after the offlining,
1805 * thus we can determine that we need to clear the node from
1806 * node_states[N_NORMAL_MEMORY].
1808 for (zt = 0; zt <= ZONE_NORMAL; zt++)
1809 present_pages += pgdat->node_zones[zt].present_pages;
1810 if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages)
1811 arg->status_change_nid_normal = zone_to_nid(zone);
1814 * We have accounted the pages from [0..ZONE_NORMAL); ZONE_HIGHMEM
1815 * does not apply as we don't support 32bit.
1816 * Here we count the possible pages from ZONE_MOVABLE.
1817 * If after having accounted all the pages, we see that the nr_pages
1818 * to be offlined is over or equal to the accounted pages,
1819 * we know that the node will become empty, and so, we can clear
1820 * it for N_MEMORY as well.
1822 present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages;
1824 if (nr_pages >= present_pages)
1825 arg->status_change_nid = zone_to_nid(zone);
1828 static void node_states_clear_node(int node, struct memory_notify *arg)
1830 if (arg->status_change_nid_normal >= 0)
1831 node_clear_state(node, N_NORMAL_MEMORY);
1833 if (arg->status_change_nid >= 0)
1834 node_clear_state(node, N_MEMORY);
1837 static int count_system_ram_pages_cb(unsigned long start_pfn,
1838 unsigned long nr_pages, void *data)
1840 unsigned long *nr_system_ram_pages = data;
1842 *nr_system_ram_pages += nr_pages;
1846 int __ref offline_pages(unsigned long start_pfn, unsigned long nr_pages,
1847 struct memory_group *group)
1849 const unsigned long end_pfn = start_pfn + nr_pages;
1850 unsigned long pfn, system_ram_pages = 0;
1851 unsigned long flags;
1853 struct memory_notify arg;
1858 * {on,off}lining is constrained to full memory sections (or more
1859 * precisely to memory blocks from the user space POV).
1860 * memmap_on_memory is an exception because it reserves initial part
1861 * of the physical memory space for vmemmaps. That space is pageblock
1864 if (WARN_ON_ONCE(!nr_pages ||
1865 !IS_ALIGNED(start_pfn, pageblock_nr_pages) ||
1866 !IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION)))
1869 mem_hotplug_begin();
1872 * Don't allow to offline memory blocks that contain holes.
1873 * Consequently, memory blocks with holes can never get onlined
1874 * via the hotplug path - online_pages() - as hotplugged memory has
1875 * no holes. This way, we e.g., don't have to worry about marking
1876 * memory holes PG_reserved, don't need pfn_valid() checks, and can
1877 * avoid using walk_system_ram_range() later.
1879 walk_system_ram_range(start_pfn, nr_pages, &system_ram_pages,
1880 count_system_ram_pages_cb);
1881 if (system_ram_pages != nr_pages) {
1883 reason = "memory holes";
1884 goto failed_removal;
1887 /* This makes hotplug much easier...and readable.
1888 we assume this for now. .*/
1889 zone = test_pages_in_a_zone(start_pfn, end_pfn);
1892 reason = "multizone range";
1893 goto failed_removal;
1895 node = zone_to_nid(zone);
1898 * Disable pcplists so that page isolation cannot race with freeing
1899 * in a way that pages from isolated pageblock are left on pcplists.
1901 zone_pcp_disable(zone);
1902 lru_cache_disable();
1904 /* set above range as isolated */
1905 ret = start_isolate_page_range(start_pfn, end_pfn,
1907 MEMORY_OFFLINE | REPORT_FAILURE);
1909 reason = "failure to isolate range";
1910 goto failed_removal_pcplists_disabled;
1913 arg.start_pfn = start_pfn;
1914 arg.nr_pages = nr_pages;
1915 node_states_check_changes_offline(nr_pages, zone, &arg);
1917 ret = memory_notify(MEM_GOING_OFFLINE, &arg);
1918 ret = notifier_to_errno(ret);
1920 reason = "notifier failure";
1921 goto failed_removal_isolated;
1927 if (signal_pending(current)) {
1929 reason = "signal backoff";
1930 goto failed_removal_isolated;
1935 ret = scan_movable_pages(pfn, end_pfn, &pfn);
1938 * TODO: fatal migration failures should bail
1941 do_migrate_range(pfn, end_pfn);
1945 if (ret != -ENOENT) {
1946 reason = "unmovable page";
1947 goto failed_removal_isolated;
1951 * Dissolve free hugepages in the memory block before doing
1952 * offlining actually in order to make hugetlbfs's object
1953 * counting consistent.
1955 ret = dissolve_free_huge_pages(start_pfn, end_pfn);
1957 reason = "failure to dissolve huge pages";
1958 goto failed_removal_isolated;
1961 ret = test_pages_isolated(start_pfn, end_pfn, MEMORY_OFFLINE);
1965 /* Mark all sections offline and remove free pages from the buddy. */
1966 __offline_isolated_pages(start_pfn, end_pfn);
1967 pr_debug("Offlined Pages %ld\n", nr_pages);
1970 * The memory sections are marked offline, and the pageblock flags
1971 * effectively stale; nobody should be touching them. Fixup the number
1972 * of isolated pageblocks, memory onlining will properly revert this.
1974 spin_lock_irqsave(&zone->lock, flags);
1975 zone->nr_isolate_pageblock -= nr_pages / pageblock_nr_pages;
1976 spin_unlock_irqrestore(&zone->lock, flags);
1979 zone_pcp_enable(zone);
1981 /* removal success */
1982 adjust_managed_page_count(pfn_to_page(start_pfn), -nr_pages);
1983 adjust_present_page_count(pfn_to_page(start_pfn), group, -nr_pages);
1985 /* reinitialise watermarks and update pcp limits */
1986 init_per_zone_wmark_min();
1988 if (!populated_zone(zone)) {
1989 zone_pcp_reset(zone);
1990 build_all_zonelists(NULL);
1993 node_states_clear_node(node, &arg);
1994 if (arg.status_change_nid >= 0) {
1996 kcompactd_stop(node);
1999 writeback_set_ratelimit();
2001 memory_notify(MEM_OFFLINE, &arg);
2002 remove_pfn_range_from_zone(zone, start_pfn, nr_pages);
2006 failed_removal_isolated:
2007 undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
2008 memory_notify(MEM_CANCEL_OFFLINE, &arg);
2009 failed_removal_pcplists_disabled:
2011 zone_pcp_enable(zone);
2013 pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n",
2014 (unsigned long long) start_pfn << PAGE_SHIFT,
2015 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1,
2017 /* pushback to free area */
2022 static int check_memblock_offlined_cb(struct memory_block *mem, void *arg)
2024 int ret = !is_memblock_offlined(mem);
2028 if (unlikely(ret)) {
2029 phys_addr_t beginpa, endpa;
2031 beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr));
2032 endpa = beginpa + memory_block_size_bytes() - 1;
2033 pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n",
2041 static int get_nr_vmemmap_pages_cb(struct memory_block *mem, void *arg)
2044 * If not set, continue with the next block.
2046 return mem->nr_vmemmap_pages;
2049 static int check_cpu_on_node(pg_data_t *pgdat)
2053 for_each_present_cpu(cpu) {
2054 if (cpu_to_node(cpu) == pgdat->node_id)
2056 * the cpu on this node isn't removed, and we can't
2057 * offline this node.
2065 static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg)
2067 int nid = *(int *)arg;
2070 * If a memory block belongs to multiple nodes, the stored nid is not
2071 * reliable. However, such blocks are always online (e.g., cannot get
2072 * offlined) and, therefore, are still spanned by the node.
2074 return mem->nid == nid ? -EEXIST : 0;
2081 * Offline a node if all memory sections and cpus of the node are removed.
2083 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2084 * and online/offline operations before this call.
2086 void try_offline_node(int nid)
2088 pg_data_t *pgdat = NODE_DATA(nid);
2092 * If the node still spans pages (especially ZONE_DEVICE), don't
2093 * offline it. A node spans memory after move_pfn_range_to_zone(),
2094 * e.g., after the memory block was onlined.
2096 if (pgdat->node_spanned_pages)
2100 * Especially offline memory blocks might not be spanned by the
2101 * node. They will get spanned by the node once they get onlined.
2102 * However, they link to the node in sysfs and can get onlined later.
2104 rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb);
2108 if (check_cpu_on_node(pgdat))
2112 * all memory/cpu of this node are removed, we can offline this
2115 node_set_offline(nid);
2116 unregister_one_node(nid);
2118 EXPORT_SYMBOL(try_offline_node);
2120 static int __ref try_remove_memory(u64 start, u64 size)
2122 struct vmem_altmap mhp_altmap = {};
2123 struct vmem_altmap *altmap = NULL;
2124 unsigned long nr_vmemmap_pages;
2125 int rc = 0, nid = NUMA_NO_NODE;
2127 BUG_ON(check_hotplug_memory_range(start, size));
2130 * All memory blocks must be offlined before removing memory. Check
2131 * whether all memory blocks in question are offline and return error
2132 * if this is not the case.
2134 * While at it, determine the nid. Note that if we'd have mixed nodes,
2135 * we'd only try to offline the last determined one -- which is good
2136 * enough for the cases we care about.
2138 rc = walk_memory_blocks(start, size, &nid, check_memblock_offlined_cb);
2143 * We only support removing memory added with MHP_MEMMAP_ON_MEMORY in
2144 * the same granularity it was added - a single memory block.
2146 if (memmap_on_memory) {
2147 nr_vmemmap_pages = walk_memory_blocks(start, size, NULL,
2148 get_nr_vmemmap_pages_cb);
2149 if (nr_vmemmap_pages) {
2150 if (size != memory_block_size_bytes()) {
2151 pr_warn("Refuse to remove %#llx - %#llx,"
2152 "wrong granularity\n",
2153 start, start + size);
2158 * Let remove_pmd_table->free_hugepage_table do the
2159 * right thing if we used vmem_altmap when hot-adding
2162 mhp_altmap.alloc = nr_vmemmap_pages;
2163 altmap = &mhp_altmap;
2167 /* remove memmap entry */
2168 firmware_map_remove(start, start + size, "System RAM");
2171 * Memory block device removal under the device_hotplug_lock is
2172 * a barrier against racing online attempts.
2174 remove_memory_block_devices(start, size);
2176 mem_hotplug_begin();
2178 arch_remove_memory(start, size, altmap);
2180 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
2181 memblock_phys_free(start, size);
2182 memblock_remove(start, size);
2185 release_mem_region_adjustable(start, size);
2187 if (nid != NUMA_NO_NODE)
2188 try_offline_node(nid);
2195 * __remove_memory - Remove memory if every memory block is offline
2196 * @start: physical address of the region to remove
2197 * @size: size of the region to remove
2199 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2200 * and online/offline operations before this call, as required by
2201 * try_offline_node().
2203 void __remove_memory(u64 start, u64 size)
2207 * trigger BUG() if some memory is not offlined prior to calling this
2210 if (try_remove_memory(start, size))
2215 * Remove memory if every memory block is offline, otherwise return -EBUSY is
2216 * some memory is not offline
2218 int remove_memory(u64 start, u64 size)
2222 lock_device_hotplug();
2223 rc = try_remove_memory(start, size);
2224 unlock_device_hotplug();
2228 EXPORT_SYMBOL_GPL(remove_memory);
2230 static int try_offline_memory_block(struct memory_block *mem, void *arg)
2232 uint8_t online_type = MMOP_ONLINE_KERNEL;
2233 uint8_t **online_types = arg;
2238 * Sense the online_type via the zone of the memory block. Offlining
2239 * with multiple zones within one memory block will be rejected
2240 * by offlining code ... so we don't care about that.
2242 page = pfn_to_online_page(section_nr_to_pfn(mem->start_section_nr));
2243 if (page && zone_idx(page_zone(page)) == ZONE_MOVABLE)
2244 online_type = MMOP_ONLINE_MOVABLE;
2246 rc = device_offline(&mem->dev);
2248 * Default is MMOP_OFFLINE - change it only if offlining succeeded,
2249 * so try_reonline_memory_block() can do the right thing.
2252 **online_types = online_type;
2255 /* Ignore if already offline. */
2256 return rc < 0 ? rc : 0;
2259 static int try_reonline_memory_block(struct memory_block *mem, void *arg)
2261 uint8_t **online_types = arg;
2264 if (**online_types != MMOP_OFFLINE) {
2265 mem->online_type = **online_types;
2266 rc = device_online(&mem->dev);
2268 pr_warn("%s: Failed to re-online memory: %d",
2272 /* Continue processing all remaining memory blocks. */
2278 * Try to offline and remove memory. Might take a long time to finish in case
2279 * memory is still in use. Primarily useful for memory devices that logically
2280 * unplugged all memory (so it's no longer in use) and want to offline + remove
2283 int offline_and_remove_memory(u64 start, u64 size)
2285 const unsigned long mb_count = size / memory_block_size_bytes();
2286 uint8_t *online_types, *tmp;
2289 if (!IS_ALIGNED(start, memory_block_size_bytes()) ||
2290 !IS_ALIGNED(size, memory_block_size_bytes()) || !size)
2294 * We'll remember the old online type of each memory block, so we can
2295 * try to revert whatever we did when offlining one memory block fails
2296 * after offlining some others succeeded.
2298 online_types = kmalloc_array(mb_count, sizeof(*online_types),
2303 * Initialize all states to MMOP_OFFLINE, so when we abort processing in
2304 * try_offline_memory_block(), we'll skip all unprocessed blocks in
2305 * try_reonline_memory_block().
2307 memset(online_types, MMOP_OFFLINE, mb_count);
2309 lock_device_hotplug();
2312 rc = walk_memory_blocks(start, size, &tmp, try_offline_memory_block);
2315 * In case we succeeded to offline all memory, remove it.
2316 * This cannot fail as it cannot get onlined in the meantime.
2319 rc = try_remove_memory(start, size);
2321 pr_err("%s: Failed to remove memory: %d", __func__, rc);
2325 * Rollback what we did. While memory onlining might theoretically fail
2326 * (nacked by a notifier), it barely ever happens.
2330 walk_memory_blocks(start, size, &tmp,
2331 try_reonline_memory_block);
2333 unlock_device_hotplug();
2335 kfree(online_types);
2338 EXPORT_SYMBOL_GPL(offline_and_remove_memory);
2339 #endif /* CONFIG_MEMORY_HOTREMOVE */